K-Means: 중심점, 할당 및 갱신 단계

K-Means is an unsupervised algorithm that partitions n data points into k non-overlapping clusters. Unlike supervised learning, there are no labels — the algorithm discovers structure purely from the feature values. K-Means is fast, scalable, and widely used for customer segmentation, image compression, and anomaly detection.

K-Means repeats three steps until convergence: 1) Initialise — randomly place k centroids in feature space. 2) Assignment — assign every point to the nearest centroid. 3) Update — move each centroid to the mean of its assigned points. The loop stops when assignments no longer change.

In each assignment step, the Euclidean distance from every point to every centroid is computed. A point is assigned to the centroid with the smallest distance. For a point x and centroid c, the squared distance is sum((x_i - c_i)^2). Using squared distance avoids the expensive square-root and gives the same ordering.

After assignment, each centroid is relocated to the arithmetic mean of all points currently in its cluster. If a cluster becomes empty (no points assigned), the centroid is usually re-initialised randomly or removed. This mean-shift minimises the total within-cluster sum of squares (WCSS) — also called inertia.

Consider four 1D points: 1, 2, 8, 9 and k=2. Init: centroids = [1, 8]. Iter 1 assignment: 1→C0, 2→C0, 8→C1, 9→C1. Iter 1 update: C0=1.5, C1=8.5. Iter 2 assignment: unchanged. Converged in 2 iterations! In higher dimensions convergence may take more steps, but the logic is identical.

Inertia (WCSS) is the sum of squared distances between each point and its cluster centroid. Lower inertia means tighter, more compact clusters. K-Means minimises inertia at each update step, but the algorithm is not guaranteed to find the global minimum — it can get stuck in local optima depending on initialisation.

Random centroid initialisation often leads to slow convergence or poor local optima. K-Means++ (the scikit-learn default via init='k-means++') seeds centroids more intelligently: the first centroid is chosen randomly, and each subsequent centroid is selected with probability proportional to its squared distance from the nearest already-chosen centroid. This spreads starting points and consistently finds better solutions.

Plotting cluster assignments on a 2D scatter shows the Voronoi partition — the decision boundaries where each point's nearest centroid changes. Plotting centroids as large stars and colouring points by cluster label makes convergence intuitive. This visualisation also reveals when clusters overlap or have unequal sizes.

Because K-Means can converge to local optima, scikit-learn runs the algorithm n_init times with different random seeds and keeps the result with the lowest inertia. The default is n_init=10. For small datasets 10 is usually sufficient; for large or tricky datasets you may raise it to 20 or 50. Always check the final inertia against the best-run inertia to diagnose poor convergence.

K-Means has several well-known weaknesses: 1) Assumes spherical clusters — it struggles with elongated or crescent shapes. 2) Sensitive to outliers — a distant outlier pulls the centroid away from the true cluster mean. 3) Requires k upfront — you must know or estimate the number of clusters before fitting. 4) Feature scale matters — always standardise features before running K-Means so large-scale variables do not dominate distances.

In practice, using sklearn.cluster.KMeans is the standard approach. Key parameters: n_clusters (k), init (default 'k-means++'), n_init, max_iter (default 300), and random_state. After fitting, km.labels_ contains cluster assignments, km.cluster_centers_ holds centroid positions, and km.inertia_ reports WCSS.

Test your understanding of K-Means clustering concepts from this lesson.

In this lesson you learned: K-Means iterates assignment and update steps until cluster memberships stabilise, inertia (WCSS) measures compactness and is minimised by each update, and K-Means++ initialisation and multiple restarts help avoid poor local optima. Next up we explore how to choose the right value of k using the elbow method and silhouette score.